Available online at www.sciencedirect.com Energy Procedia 00 (2013) 000 000 www.elsevier.com/locate/procedia SiliconPV: March 25-27, 2013, Hamelin, Germany Solar cell performance prediction using advanced analysis methods on optical images of as-cut wafers Marko Turek*, Dominik Lausch Fraunhofer Center for Silicon Photovoltaics CSP, W.-Hülse-Str. 1, 06120 Halle (Saale) Abstract A quantitative evaluation of the material quality of as-cut wafers with respect to the corresponding solar cell performance is the basis for a reliable quality control. A number of techniques have been recently developed with most of them using photoluminescence (PL) images as a starting point for the application of various image processing methods. In this work, a new empirical approach is demonstrated that relies on the analysis of optical images. We investigate both optical and PL-images of as-cut wafers using advanced image processing algorithms and compare their predictive power when applied to as-cut wafers. While the optical images of as-cut wafers are much easier acquired, our results show that they nevertheless can be used for a quantitative rating that correlates with the electrical properties of the processed cells. 2013 The Authors. Published by Elsevier Ltd. Selection and/or peer-review under responsibility of the scientific committee of the SiliconPV 2013 conference Keywords: Type your keywords here, separated by semicolons ; 1. Introduction A reliable quality control is an important contribution for increasing the cost efficiency in the very competitive photovoltaic industry. On the level of as-cut wafers, quality ratings based on PL-images have been investigated intensively during the last few years. Image processing algorithms have been developed that allow the quantitative determination of edge regions and the extraction of quantities that are related to the density of electrically active dislocations. [1, 2, 3, 4, 5, 6] The effective sizes of the edge regions and effective dislocation densities obtained from these image processing approaches were shown to correlate * Marko Turek. Tel.: +49 345 5589-414; fax: +49 345 5589-101. E-mail address: marko.turek@csp.fraunhofer.de.
with characteristic cell data like the open circuit voltage. Another approach relying on optical images of etched wafers after saw damage removal has also been shown to yield a characteristic wafer parameter that correlates with the open circuit voltage of the corresponding cells. [7] In this work, we will present new results that are based on the analysis not only of PL-images but also of optically obtained images. All spatially resolved data sets are obtained from as-cut wafers without any further sample preparation like acidic etching. A number of image processing algorithms are applied to either spatially resolved data set leading to several effective parameters which can be correlated to the corresponding cell data. In particular, the photoluminescence-images (PL) are used to identify the fraction of edge regions (ER) and the fraction of high-quality regions (HQ). The optical images (OPT) are processed such that grain boundaries can be identified. Effective parameters that are related to the grain boundary length (GBL) and grain size (GS) are extracted. Additionally, our investigation goes beyond previously published works by extending the correlation analysis to the short circuit current I sc and the open circuit voltage U oc separately. It is shown how the parameters obtained by the image processing correlate with the cell data U oc and I sc individually. The relationship with the cell efficiency is then established in a straightforward manner as we show that the solar cell fill factors are rather constant for all cells under consideration due to a mixed and simultaneous cell processing. 2. Image generation and processing Our analysis is based on an empirical approach using data that can be obtained from as-cut wafers, i.e. PL-imaging or optical imaging. The methods are applied to wafers taken from two multi-crystalline Silicon ingots. The calculated quantities based on the image processing algorithms, i.e. ER (edge region size PL), HQ (high quality area size PL), GBL (grain boundary length OPT), and GS (grain size OPT), where then compared to typical cell data, i.e. I sc (short circuit current) and U oc (open circuit voltage), which were measured on cells coming from the direct vicinity of the individual as-cuts wafers, see Figure 1. Figure 1: Data analysis scheme: Short circuit current I sc and open circuit voltage V oc are obtained from a cell; grain boundary length GBL and grain size GS are obtained from optical imaging while edge region size ER and high quality area size HQ are determined from PL imaging. 2.1. Experimental setup Wafers and solar cells from two mc-si blocks are analyzed: block B1 coming from a corner of an ingot and block B2 coming from an edge side. Around 15 wafers equally distributed over the block height are investigated for both blocks. The resistivity of all wafers was measured and PL-images of the wafers were taken using a BTI-LiS-R1 tool. Optical images of the wafers were obtained by scanning them with a standard commercially available optical scanner. Alternatively, the optical images could be obtained using an inline IR-inspection tool as it is employed in most wafering and cell production lines. While the
PL-imaging requires a rather expensive and elaborate imaging setup the optical images can thus easily be generated. In a further step, the neighboring wafers were processed to cells using a standard cell process and the characteristic cell parameters, i.e. short circuit current I sc and open circuit voltage U oc, were determined. 2.2. Image processing and data analysis The PL- and optical images are analyzed using different image processing algorithms. First, all edge regions are identified based on the PL-image using an adaptive threshold algorithm. This algorithm separates regions within the PL image that are close to the edges and are characterized by a significantly lower PL signal than the average PL-value. The fraction of the identified edge region compared to the total wafer area is described by the parameter ER, see Figure 2 (middle). The remaining center region is then analyzed by a second adaptive threshold algorithm which determines the fraction HQ of the area that is dominated by a low recombination activity, see Figure 2 (right). This algorithm is independent of the average PL-signal and therefor of the doping level. The identified HQ-areas are typically related to a lower density of structural defects. However, a strict relationship between structural defects and electrically active defects is not always given since not all structural defects are electrically active. [8] Figure 2: PL-image processing scheme for an as-cut wafer: first the edge regions are identified yielding the parameter ER and then the high quality areas are identified within the inner region (black pixels in right-most image) giving parameter HQ. The multi-grain structure of mc-si wafers is reflected by regions of different grey values in the optical images. The boundaries of these regions can thus be identified using an image processing algorithm based on the evaluation of local gradients. The result is shown in the middle part of Figure 3 where an overlayimage of the original OPT-data and the identified edges (shown in red) is presented. It yields a characteristic quantity GBL that is related to the length of the grain boundaries. Based on these boundary lines, a second algorithm identifies topologically connected areas, see Figure 3 (right-most image). The distribution of the size of all identified connected areas is investigated and a key parameter GS extracted that describes the typical area size. Other parameters, e.g. the area scale describing the exponentially decaying size distribution, were also extracted and their correlation with the cell characteristics will be discussed elsewhere.
Figure 3: Optical image processing scheme for an as-cut wafer: first, lines with large gradients in the grey-value are identified giving parameter GBL; then topologically connected areas are determined and analysed with respect to their size yielding GS. 3. Correlation of image analysis and solar cell properties In either case, PL-imaging and OPT-imaging, the image processing aims at a data analysis that focusses on the structural properties of the images rather than on a simple statistical analysis. These properties are then described by single, quantifiable effective wafer parameters which allow a systematic quantitative correlation of the properties of as-cut wafers and the corresponding solar cells. 3.1. Samples under investigation and solar cell characterization The image processing and parameter extraction has been performed using a software algorithm that is capable of automatically analyzing the PL- and OPT-data of many wafers. The effective parameters ER (representing the size of the edge region and obtained from PL-images), HQ (being characteristic for the relative size of the high-quality region, extracted from PL-images), GBL (corresponding to the grain boundary length, determined from optical images), and GS (describing the average grain size) were compared to the cell parameters I sc and U oc. In order to obtain conclusive results, two ingots were chosen that exhibit a very characteristic I sc - and U oc -profile, see Figure 4. While block B1 exhibits a rather continuous decrease of the short circuit current when going from bottom to top the open circuit voltage shows a rather prominent minimum at around two thirds of the block height. The wafers and cells coming from the very bottom are an exception due to the contaminations from the crucible. In contrast, Block B2 shows for both quantities a maximum after around one third of the block height, see Figure 4 (right). Figure 4: Height profile of open circuit voltage U oc and short circuit current I sc for block B1 (left) and block B2 (right).
In order to exclude any specific doping or cell process related factors, the resistivity of the wafers and the fill factor of the cells have been analyzed. It is important to ensure that there is not cell process related performance variation when studying the correlation between the material property of the as-cut wafers and the solar cell performance. A rather sensitive quantity for the analysis of the cell process is given by the fill factor. It is found that the fill factor remains rather constant over the entire block height while the resistivity shows the typical decaying trend from bottom to top of block B1. Block B2 shows a similar result except that there is a weak trend towards smaller fill factors with increasing block height. The reason for this fill factor behavior was found to lie in a weakly decreasing shunt resistance. The result of this analysis is that neither doping nor cell processing can be a reason for the observed non-monotonic behavior of the cell characteristics shown in Figure 4. Furthermore, it has been verified that our PL-algorithm is capable of predicting solar cell parameters when being applied to solar cells directly instead of as-cut wafers. This test of the applicability of the PLapproach is essential when it is to be compared with an approach based on optical images. As a result, we find that in the lower third of block B2 there is a significant influence of the edge region described by parameter ER, see Figure 5 (left). These edge regions on cell level (which are quantified by our algorithm as parameter ER) decrease in size with increasing block height and are negligible for the upper two thirds of the block. In the lower third of the block they imply a reduction of the short circuit current which thus reaches a maximum at around one third of the block height, see Figure 5 (left). Beyond that point, other loss mechanisms, which are caused by recombination active defects, start to dominate leading to a decreasing short circuit current again. These loss mechanisms are characterized by the relative size of the high quality area described by parameter HQ, see Figure 5 (right). This parameter obtained by our PLalgorithm to cells shows a very good correlation with the open circuit voltage which has been shown by other groups using similar algorithms before. [9] Thus, it can be concluded that the U oc and I sc characteristics of block B2 with the maximum at around one third of the block height are caused by the decreasing influence of the edge area and the increasing influence of the defects that are spread over the entire cell. Figure 5: Block B2 height profile of short circuit current and edge area fraction ER (left); height profile of open circuit voltage and high quality area HQ (right). Having checked that our PL-algorithm works well on cell level, we have applied it to as-cut wafers in order to compare its predictive power with results one can obtain from algorithms that are based on optical images.
3.2. Predictive power of photoluminescence imaging on as-cut wafers When analyzing PL-images it is important to apply image processing algorithms that are independent of the overall trend which is given by the PL-signal averaged over the entire wafer as this average is mostly determined by the doping profile within the block. This correlation can be seen in Figure 6. Hence, it becomes clear that simple quantities like the average PL-signal cannot be expected to yield correlations to the U oc and I sc profiles shown in Figure 4. Besides the doping concentration, the effective minority carrier lifetime is the second important contribution to the PL signal. This is reflected, for example, by the large deviations between the PL signal and the doping concentration at the lower parts of the ingots, see Figure 6. It has to be noted, that the measurable effective lifetime correlates to the bulk lifetime only for very low lifetimes due to the small thickness and the large surface recombination of the as-cut wafers. [10] Hence, the PL-signal variation caused by the bulk lifetime of as-cut wafers is rather low. Nevertheless, it is the bulk lifetime and not the measurable effective lifetime that influences the cell efficiency. Figure 6: Height profile of doping concentration (calculated from resistivity measurements) and average PL-signal for block B1 (left) and block B2 (right). A more advanced approach for the evaluation of spatially resolved PL-measurements is based on the detection of edge areas within the ingot that were close to the crucible. The height profile of this edge area size of as-cut wafers (parameter ER) is compared to the short circuit current as shown in Figure 7. When analyzing the results for block B2, it becomes clear that the edge regions on wafer level, see Figure 7 (right), are significantly different from the edge regions found on cell level, see Figure 5 (left). This is caused by the cell process that effectively reduces these edge regions. It furthermore implies that the edge regions determined on as-cut wafers by PL-imaging do not have a significant predictive power for the cell parameters. In particular, there is no correlation between the disappearance of the edge regions on wafer level and the maximum found for the short circuit current. For block B1, the size of the edge regions on wafer level decreases continuously with the ingot height. This would indicate an increasing short circuit current, however the opposite trend is observed for block B1, see Figure 7 (left). Hence, it can be concluded that the edge regions that are detected by the PL-algorithm on wafer level are more or less completely insignificant on cell level where other loss mechanisms like recombination active defects that are spread over the entire wafer area dominate.
Figure 7: Height profile of short circuit current and edge area fraction ER obtained from PL-image analysis for block B1 (left) and block B2 (right). Secondly, the relation between the quantity HQ (obtained by PL-imaging and being qualitatively related to high quality areas) and the cell parameters is investigated. The result is presented in Figure 8 and shows that in case of block B1 the extracted parameter HQ follows closely the open circuit voltage U oc. This rather close correlation holds true for block B1 while it does not for block B2. The second block exhibits an U oc HQ correlation for the lower third of the block only. Again, this discrepancy between the HQ-parameter determined by PL-imaging on as-cut wafer level, see Figure 8 (right), and the same quantity on cell level, see Figure 5 (right), is caused by the cell process which activates certain defect types. [8] Whether this effect occurs, like for block B2, or does not occur, like for block B1, depends on the material itself. Thus, in general a high predictive power of a PL-based image analysis is not always granted and depends on the material and its prevalent defect types. Figure 8: Height profile of open circuit voltage and high quality area fraction HQ obtained from PL-image analysis for block B1 (left) and block B2 (right). In the final analysis, we compare these findings that were obtained on as-cut wafers using PL-imaging with our new approach based on the technically much simpler optical imaging. 3.3. Evaluation of effective wafer parameters obtained by optical imaging of as-cut wafers Our analysis of the optical images is based on the effective parameter GS which is related to a typical grain size obtained from the distribution of all grain sizes within a wafer. The correlation of the typical grain size GS obtained by an optical imaging with the open circuit voltage U oc is shown in Figure 9.
Block B1 shows a clear correlation between U oc and GS, in particular both quantities show a minimum at the same position within the ingot. At the very bottom of the block this correlation cannot be established since the ingot contaminations due to the crucible are not directly reflected in the optically visible material structure. Excluding this region at the very bottom of the ingot, the result implies that the grains are not only a structural property but also influence the electrical cell properties in a rather direct manner. On the other hand, a similar correlation can be found for block B2 within the lowest third of the block only. At higher block positions, the extracted typical grain size GS remains rather constant while some other material effect appears to dominate the electrical properties characterized by U oc. When studying the correlation between the inverse grain boundary length 1/GBL and the open circuit voltage a very similar result can be found. This is intuitively clear as the grain size and the grain boundary length are strongly correlated as long as the grains do not exhibit any special geometric shapes. Comparing these findings based on optical imaging, see Figure 9, with the results obtained from PLimaging, see Figure 8, one finds that our novel approach using the optical images has the same predictive power as the PL-imaging does. Blocks for which the PL-approach yields reasonable correlations, i.e. block B1, exhibit an equally good correlation when applying optical imaging. On the other hand, blocks where the optical approach fails to predict the cell parameters, i.e. block B2, show the same lacking correlation for the PL-approach. Figure 9: Height profile of open circuit voltage and grain size indicator GS for block B1 (left) and block B2 (right). 4. Conclusions A new approach for a solar cell performance prediction based on advanced image processing algorithms for as-cut wafers has been investigated. To this end, two different types of spatially resolved data sets have been analyzed: images obtained by photoluminescence measurements and optical images of as-cut wafers. The optical images are obtained with much less technical effort than the photoluminescence images. Nevertheless are capable of yielding similar information related to the cell performance although they do not yield any direct information on the electrical material properties. In particular, we found a correlation between the optically detected grain size and the open circuit voltage which is of equal predictive power than an approach based on photoluminescence imaging. In contrast to the optical imaging approach, the photoluminescence images of as-cut wafers allow the determination of edge regions that are contaminated by the crucible. However, these regions are strongly changed during the cell process such that the information about the size of the edge regions obtained by photoluminescence measurements does not allow a prediction of the cell performance. While some
material types can be characterized rather well predicting cell parameters both approaches photoluminescence imaging and optical imaging fail in a similar way for other material types. This is caused by the cell process that is applied to the as-cut wafers and causes the activation of certain defect types as well as the reduction of the contaminated edge regions that were close to the crucible. Acknowledgements The authors acknowledge the funding of the Ministerium für Wirtschaft und Arbeit des Landes Sachsen-Anhalt and the European Union within the MiniSZ-project (FuE 56/10). Valuable discussions with Christian Hagendorf and the measurement support by Jan Lich are gratefully acknowledged. Part of the material was kindly provided by Kai Petter which is also gratefully acknowledged. References [1] M. Demant, M. Glatthaar, J. Haunschild, and S. Rein. Analysis of luminescence images applying pattern recognition techniques. In Proceedings of the 25th EU PVSEC, Valencia, Spain, 2010. [2] Jonas Haunschild, Markus Glatthaar, Matthias Demant, Jan Nievendick, Markus Motzko, Stefan Rein, and Eicke R. Weber. Quality control of as-cut multicrystalline silicon wafers using photoluminescence imaging for solar cell production. Solar Energy Materials and Solar Cells, 94(12):2007 2012, 2010. [3] W. McMillan, T. Trupke, J.W. Weber, M. Wagner, U. Mareck, Y.C. Chou, and J. Wong. In-line monitoring of electrical wafer quality using photoluminescence imaging. In Proceedings of the 25th EU PVSEC, Valencia, Spain, 2010. [4] B. Birkmann, A. Hüsler, A. Seidl, K. Ramspeck, and H. Nagel. Analysis of multicrystalline wafers originating from corner and edge bricks and forecast of cell properties. In Proceedings of the 26th EU PVSEC, Hamburg, Germany, 2011. [5] B. True, A. Stavrides, and I. Latchford. Image processing techniques for correlation of photoluminescence images of ascut wafers with final cell iv parameters. In Proceedings of the 26th EU PVSEC, Hamburg, Germany, 2011. [6] Ronald A. Sinton, Jonas Haunschild, Matthias Demant, and Stefan Rein. Comparing lifetime and photoluminescence imaging pattern recognition methodologies for predicting solar cell results based on as-cut wafer properties. Progress in Photovoltaics: Research and Applications, 2012. [7] R. Bakowskie, G. Kesser, R. Richter, D. Lausch, A. Eidner, P. Clemens, and K Petter. Fast method to determine the structural defect density of 156 x 156 mm^2 mc-si wafers. Energy Procedia, 27:179, 2012. [8] D. Lausch, K. Petter, R. Bakowskie, J. Bauer, O. Breitenstein, and Ch. Hagendorf. Classification and investigation of recombination-active defect structures in multicrystalline silicon solar cells - recombination models. In Proceedings of the 27th EU PVSEC, Frankfurt a.m., Germany, 2012. [9] R. Bakowskie, R. Lantzsch, T. Kaden, K.G. Eller, D. Lausch, Y. Ludwig, and K. Petter. Comparison of recombination active defects in multicrystalline silicon by means of pl imaging and reverse biased electroluminescence. In Proceedings of the 26th EU PVSEC, Hamburg, Germany, 2011. [10] M. Turek. Interplay of bulk and surface properties for steady-state measurements of minority carrier lifetimes. J. Appl. Phys., 111:123703, 2012.